Climate Foundation Models Face Robustness Challenges Under No-Analog Shifts
A recent study published on arXiv (2603.23043) examines the resilience of climate foundation models in the face of no-analog distribution shifts. The rapid advancement of climate change presents non-stationarities that hinder the capacity of machine learning-based climate emulators to extend beyond their training distributions. While these emulators serve as computationally efficient substitutes for conventional Earth System Models, their dependability becomes a concern under future climate conditions lacking historical precedents. A significant challenge is data contamination, as many models are developed using simulations that incorporate future scenarios, obscuring their genuine out-of-distribution performance. The research evaluates the OOD robustness of three architectures—U-Net, ConvLSTM, and ClimaX—by limiting them to historical data.
Key facts
- Study assesses robustness of climate foundation models under no-analog distribution shifts
- Climate change introduces non-stationarities challenging ML emulators
- Emulators are efficient alternatives to Earth System Models
- Reliability is a bottleneck under no-analog future climate states
- Data contamination masks true out-of-distribution performance
- Benchmarks three architectures: U-Net, ConvLSTM, ClimaX
- Models restricted to historical data for evaluation
- Published on arXiv with ID 2603.23043
Entities
Institutions
- arXiv